Custom AI servers,
built to order
We build GPU servers around your workload: GPUs, CPU, memory, storage, power and cooling sized by engineers, assembled and tested before delivery and shipped with warranty anywhere in the EU. Configuration and quote within one business day.
One server, sized for one job
We start from the workload, not from a parts list. These are the four jobs we build for most often.
Multi-GPU nodes for model training and fine-tuning: NVIDIA H200 NVL with NVLink bridges or RTX PRO 6000 Blackwell Server Edition, fast NVMe scratch storage and high-speed networking for multi-node clusters.
Servers for private LLMs and RAG in production: 2 to 8 GPUs per node sized by model size and concurrent users, large ECC memory and NVMe storage for indexes and documents.
GPU nodes for virtual workstations and render farms: RTX PRO Blackwell, L40S or L4 with NVIDIA vGPU licensing, sized by seats or frames per card, not by the brochure.
GPU workstations and short-depth servers for labs, pilots and branch sites: Threadripper PRO or Xeon W platforms with RTX PRO Blackwell Workstation cards.
Components we size and source
- NVIDIA RTX PRO 6000 Blackwell: Server & Workstation; RTX PRO 5000: Workstation
- NVIDIA H200 NVL: 141 GB HBM3e, NVLink bridges
- NVIDIA L40S · L4
- AMD EPYC
- Intel Xeon Scalable
- AMD Ryzen Threadripper PRO
- DDR5 ECC RDIMM / LRDIMM
- MRDIMM: up to 8800 MT/s
- Sized to the GPU pool, not the minimum
- NVMe U.2 data-centre SSD
- SAS 24G enterprise SSD
- Scratch, model store and index tiers
- 25 / 100 / 200 / 400G NICs
- Cluster fabrics for multi-node training
- Out-of-band management (IPMI / BMC)
- 2U / 4U GPU chassis, redundant PSUs
- Air or liquid cooling by GPU count and site
- Rack, power and airflow check before we quote
Three reference configurations
Not a price list: the sizing conversation decides the final build. These show the scale we work at.
| CONFIGURATION | GPUs | PLATFORM | BUILT FOR |
|---|---|---|---|
| Private AI starter | 2× RTX PRO 6000 Blackwell Server Edition (2×96 GB) | Dual AMD EPYC · 384 GB DDR5 ECC · 2× NVMe | Pilots, RAG assistants, models up to ~70B parameters |
| Inference node | 4× RTX PRO 6000 Blackwell Server Edition (4×96 GB) | Dual AMD EPYC · 768 GB DDR5 ECC · 4× NVMe | Production inference for several teams, multi-model serving |
| Training node | 8× NVIDIA H200 NVL (8×141 GB, NVLink bridges) | Dual AMD EPYC · 2.3 TB DDR5 ECC (24×96 GB) · 8× NVMe · 400G | Fine-tuning and training, multi-node clusters |
Memory and cores follow NVIDIA’s certified-systems guidance (at least twice the GPU memory, at least six cores per GPU); more on request.
A server is step one
If you also need what runs on it (private LLMs, RAG on your documents, MLOps), that is Private AI/ML by the same team. NVIDIA AI Enterprise and vGPU licences come on the same invoice as the hardware.
Custom builds, in short
How long does a custom build take?
Can you build on our chassis, or with parts we already own?
What warranty do we get?
Do you install the software?
Do you deliver outside the EU?
We do not know what we need yet. Where do we start?
From our engineering blog
Weights, KV cache, overhead and concurrent users: the arithmetic behind the GPU count.
Read the article →What a multi-GPU node draws, how much heat it makes and what the rack has to deliver.
Read the article →Identical compute, 1.5× the memory and 1.22× the bandwidth: when the upgrade pays, and when an existing H100 NVL is enough.
Read the article →Request a configuration
Tell us the workload; an engineer replies with a configuration and a quote within one business day